Each row is a layer (or, for GoogLeNet and ResNet, one block): output shape, receptive field, and parameter count, with the count drawn as a bar on a log scale. The most expensive row is marked in gold — find it in each network before believing any claim about where parameters live. In build your own, the head is always global-average-pool + linear to 10 classes; watch what each added layer does to the three columns.